| Message ID | BLU436-SMTP164FFBA31AD0A827A87B35997230@phx.gbl |
|---|---|
| State | New |
| Headers | show |
On 02/13/2015 05:22 AM, John David Anglin wrote: > + /* Reload sometimes tries to put const data symbolic operands in > + readonly memory. The HP SOM linker doesn't allow symbolic data > + in readonly memory. */ > + if (TARGET_SOM > + && !function_label_operand (x, VOIDmode) > + && symbolic_operand (x, VOIDmode)) > + return true; You probably want to remove the SOM test. Even if ELF can represent this, it will lead to DT_TEXTREL and relocations against the read-only memory. r~
On 2015-02-13, at 12:08 PM, Richard Henderson wrote: > On 02/13/2015 05:22 AM, John David Anglin wrote: >> + /* Reload sometimes tries to put const data symbolic operands in >> + readonly memory. The HP SOM linker doesn't allow symbolic data >> + in readonly memory. */ >> + if (TARGET_SOM >> + && !function_label_operand (x, VOIDmode) >> + && symbolic_operand (x, VOIDmode)) >> + return true; > > You probably want to remove the SOM test. Even if ELF can represent this, it > will lead to DT_TEXTREL and relocations against the read-only memory. Unfortunately, this creates new problems. When there is a constant offset that won't fit in 14 bits, we ICE. I think this could be improved to 21 bits but it's tough to load an operand with a larger offset with just one register, particularly when generating PIC code. I need to research what happens on SOM with large offsets. Possibly the constant can somehow be forced into the data section where the relocations aren't a problem? Dave -- John David Anglin dave.anglin@bell.net
On 02/14/2015 06:50 AM, John David Anglin wrote: > Possibly the constant can somehow be forced into the data section where the relocations > aren't a problem? Hmm. It looks like we might already do that. See default_select_rtx_section. r~
On 2015-02-16, at 11:38 AM, Richard Henderson wrote: >> >> Possibly the constant can somehow be forced into the data section where the relocations >> aren't a problem? > > Hmm. It looks like we might already do that. See default_select_rtx_section. Thanks, I see the problem. default_reloc_rw_mask returns 0 when not generating PIC code, so rtx went to readonly_data_section. I was thinking that pa_select_section was somehow broken. Dave -- John David Anglin dave.anglin@bell.net
Index: config/pa/constraints.md =================================================================== --- config/pa/constraints.md (revision 220598) +++ config/pa/constraints.md (working copy) @@ -106,7 +106,7 @@ (and (match_code "mem") (match_test "IS_LO_SUM_DLT_ADDR_P (XEXP (op, 0))"))) -(define_constraint "Q" +(define_memory_constraint "Q" "A memory operand that can be used as the destination operand of an integer store, or the source operand of an integer load. That is any memory operand that isn't a symbolic, indexed or lo_sum memory @@ -122,7 +122,7 @@ (and (match_code "mem") (match_test "IS_INDEX_ADDR_P (XEXP (op, 0))"))) -(define_constraint "T" +(define_memory_constraint "T" "A memory operand for floating-point loads and stores." (match_test "floating_point_store_memory_operand (op, mode)")) Index: config/pa/pa.c =================================================================== --- config/pa/pa.c (revision 220598) +++ config/pa/pa.c (working copy) @@ -1569,6 +1569,14 @@ static bool pa_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED, rtx x) { + /* Reload sometimes tries to put const data symbolic operands in + readonly memory. The HP SOM linker doesn't allow symbolic data + in readonly memory. */ + if (TARGET_SOM + && !function_label_operand (x, VOIDmode) + && symbolic_operand (x, VOIDmode)) + return true; + return tls_referenced_p (x); }